AMCA-DTA: An Atom-level Multimodal Cross-Attention Framework for Drug-Target Affinity Prediction

Accurate prediction of Drug-Target Binding Affinity (DTA) is a pivotal task in computational drug discovery, with direct implications for virtual screening, lead optimization, and drug repositioning. While deep learning approaches have made significant strides in DTA prediction, existing methods often suffer from two key limitations: (1) reliance on single-modality representations that fail to capture the complementary nature of sequence and structural information, and (2) coarse-grained fusion strategies that cannot model the fine-grained, atom-level interactions governing molecular recognition. To address these challenges, we propose AMCA-DTA (Atom-level Multimodal Cross-Attention for DTA), a novel framework that integrates Graph Transformers for drug and protein structure encoding with a specialized atom-level cross-attention mechanism. Our model explicitly captures bidirectional dependencies between drug atoms and protein residues, enabling fine-grained cross-modal interaction learning. Comprehensive evaluations on the Davis and KIBA benchmark datasets demonstrate that AMCA-DTA achieves state-of-the-art performance, outperforming existing methods across multiple evaluation metrics including MSE, CI, and $r_m^2$. Ablation studies further validate the contribution of each key component, highlighting the importance of both structural encoding and fine-grained interaction modeling for accurate affinity prediction.